The UN puts its statistics at the service of AI agents, after a test revealed just 21% accuracy in models

By Zendoric · September 18, 2026.

The United Nations and Google have unveiled UN System Data Commons, a platform that allows the UN's official statistics to be queried in natural language and connected directly to AI agents. The project builds on Data Commons, the open-source tool Google launched in 2018 to organize public databases, and adds support for the Model Context Protocol (MCP) —the standard, originally driven by Anthropic, that lets an agent connect to external data sources instead of relying solely on what it learned during training.

The trigger for the project is an uncomfortable figure. A UNICEF benchmark, with more than 133,000 responses generated by six large language models —OpenAI's GPT-4o and GPT-4o-mini, Anthropic's Claude Sonnet 4.5 and Haiku 4.5, and Google's Gemini 2.5 Flash and Gemini 2.0 Flash— found average accuracy of barely 21.2% in answering questions about global development indicators, as João Pedro Azevedo, UNICEF's chief statistician, explained to reporters. Around three out of every five answers did not even manage to offer a usable figure: the models opted for evasion. And when the same question was put to the same models two days later, those that did give a number agreed with themselves only about half the time. The study, not yet peer-reviewed, will be published along with its methodology, code and data.

That fragility coexists with adoption that is already real: UNICEF's data portal receives more than 6 million visits a month, and those arriving from ChatGPT links grew 67% year on year between January and mid-September, to account for 6.4% of this year's sessions. Adding up all AI assistants, Azevedo calculates that they already generate close to one in every ten visits. The public, in other words, is already asking chatbots for development statistics before a reliable infrastructure exists to answer them.

UN System Data Commons aims to close that gap. Twenty-five United Nations system entities have committed to the project and nearly 20 already had data available at launch; the goal is to incorporate 80% of the UN's statistical datasets by 2027. Google.org has contributed $2 million for capacity building, and the rollout followed, according to Prem Ramaswami (head of Data Commons at Google), a "training of trainers" model: the instance is governed and administered by the UN itself, not by Google. Each record also retains its provenance, so that a user —or an agent— can trace any figure back to the agency and original source that produced it.

In a demonstration, Google showed an agent that combined several UN metrics to generate dashboards and reports automatically, and another that, starting from a question about the impact of PEPFAR (the U.S. aid program against AIDS) in Africa, located statistics on HIV infections, mortality and life expectancy and turned them into an infographic. It is a demonstration of convenience, not proof that the conclusion generated was correct. Ramaswami himself warned of this: connecting a model to reliable sources does not guarantee that its conclusions will be reliable too, because models can still misread nuances, and a person must always review the result before citing it.

Here it is worth separating what has been demonstrated from what has been promised, something we repeat at Zendoric every time a big tech company unveils infrastructure for agents: giving a model access to quality data and traceability does not solve the underlying problem, which is interpretation. MCP solves "where the data is"; it does not solve "what this figure means within this series, with this definition and this period", which is exactly where the six models in UNICEF's benchmark failed. That is why the UN insists, almost as a stylistic clause, that human review will remain mandatory: it is not rhetorical caution, it is the direct conclusion from its own data.

Even so, the project matters beyond the UN. It is one more piece in the race to build the "plumbing" of agentic AI —the standards and pipelines through which agents access the real world—: the winner is whoever controls how agents connect to information, not just whoever has the smartest model. That the chosen standard is MCP, born at Anthropic, and that the party deploying it at this institutional scale is Google, says a great deal about where the industry is converging: toward de facto interoperability between rival labs, even as they compete fiercely on everything else.

In the long run, this is exactly the kind of infrastructure we need if AI is to genuinely help solve global development problems —disease tracking, poverty, child mortality— rather than merely writing about them with invented data. A statistic with verifiable provenance, accessible in natural language, is a real building block toward the informational abundance we defend as a horizon at Zendoric. But this benchmark's 21.2% accuracy is a useful reminder that the distance between having the right data available and producing a correct analysis remains enormous, and that no one —neither the UN nor Google— claims today to have closed it.